if_link_state_change can execute the network stack that is expected to
not run in hardware interrupt (at least now), however network drivers
may call it in hardware interrupt. Avoid that by introducing a new
softint for if_link_state_change.
The original patch is provided by mlelstv@ and tweaked a bit by me.
Should fix PR kern/50602.
Thanks to introducing softint-based if_input, the entire bridge code now
never run in hardware interrupt context. So we can simplify the code.
- Remove spin mutexes
- They were needed because some code of bridge could run in
hardware interrupt context
- We now need only an adaptive mutex for each shared object
(a member list and a forwarding table)
- Remove pktqueue
- bridge_input is already in softint, using another softint
(for bridge_forward) is useless
- Packet distribution should be down at device drivers
-gcc: internal compiler error: Killed (program cc1plus)
+gcc: internal compiler error: Killed (program cc1plus received signal 9)
which was lost in an update somewhere after netbsd-6.
This clarification was found in the past to significantly reduce the
number of bug reports caused by people running out of swap, such as
(most likely) seen again today in PR 50802.
XXX: pullup-7
el_getc() for the WIDECHAR case, that is, the version in eln.c.
For a UTF-8 locale, it is broken in four ways:
1. If the character read is outside the ASCII range, the function
does an undefined cast from wchar_t to char. Even if wchar_t
is internally represented as UCS-4, that is wrong and dangerous
because characters beyond codepoint U+0255 get their high bits
truncated, meaning that perfectly valid printable Unicode
characters get mapped to arbitrary bytes, even the ASCII escape
character for some Unicode characters. But wchar_t need not
be implemented in terms of UCS-4, so the outcome of this function
is undefined for any and all input.
2. If insufficient space is available for the result, the function
fails to detect failure and returns garbage rather than -1 as
specified in the documentation.
3. The documentation says that errno will be set on failure, but
that doesn't happen either in the above case.
4. Even for ASCII characters, the results may be wrong if wchar_t
is not using UCS-4.
As we have seen before, "histedit.h" can never get rid of including
the <wchar.h> header because using the data types defined there is
deeply ingrained in the public interfaces of libedit.
Now POSIX unconditionally requires that <wchar.h> defines the type
wint_t. Consequently, it can be used unconditionally, no matter
whether WIDECHAR is active or not. Consequently, the #define Int
is pointless.
Note that removing it is not gratuitious churn. Auditing for
integer signedness problems is already hard when only fundamental
types like "int" and "unsigned" are involved. It gets very hard
when types come into the picture that have platform-dependent
signedness, like "char" and "wint_t". Adding yet another layer
on top, changing both the signedness and the width in a platform-
dependent way, makes auditing yet harder, which IMHO is really
dangerous. Note that while removing the #define, i already found
one bug caused by this excessive complication - in the function
re_putc() in refresh.c. If WIDECHAR was defined, it printed an
Int = wint_t value with %c. Fortunately, that bug only affects
debugging, not production. The fix is contained in the patch.
With WIDECHAR, this doesn't change anything. For the case without
WIDECHAR, i checked that none of the places wants to store values
that might not fit in wint_t.
This only changes internal interfaces; public ones remain unchanged.
Next step: Remove #ifdef'ing in read_char(), in the same style
as we did for setlocale(3) in el.c.
A few remarks are required to explain the choices made.
* On first sight, handling mbrtowc(3) seems a bit less trivial
than handling setlocale(3) because its prototype uses the data
type mbstate_t from <wchar.h>. However, it turns out that
"histedit.h" already includes <wchar.h> unconditionally (i don't
like headers including other headers, but that ship has sailed,
people are by now certainly used to the fact that including
"histedit.h" doesn't require including <wchar.h> before), and
"histedit.h" is of course included all over the place. So from
that perspective, there is no problem with using mbrtowc(3)
unconditionally ever for !WIDECHAR.
* However, <wchar.h> also defines the mbrtowc(3) prototype,
so we cannot just #define mbrtowc away, or including the header
will break. It would also be a bad idea to porovide a local
implementation of mbrtowc() and hope that it overrides the one
in libc. Besides, the required prototype is subtly different:
While mbrtowc(3) takes "wchar_t *" as its first argument, we
need a function that takes "Char *". So unfortunately, we have
to keep a ct_mbrtowc #define, at least until we can maybe get
rid of "Char *" in the more remote future.
* After getting rid of the #else clause in read_char(), we can
pull "return 1;" into the default: clause. After that, we can
get rid of the ugly "goto again_lastbyte;" and just "break;".
As a bonus, that also gets rid of the ugly CONSTCOND.
* While here, delete the unused ct_mbtowc() from chartype.h.
Return error code, not 0 (!), on bus_space_subregion failure.
In answer to `XXX error branch' comment: if nouveau_barobj_ctor
fails, then the caller will call nouveau_barobj_dtor too. So there's
no leak here.
Unlikely to fix any observed bugs with nouveau -- there's no error
branch in the Linux side here. But maybe it will catch some other
bug earlier.
